The Wisdom of Physics: Sovereign Coordination as a Physical Substrate for Multi-Agent Autonomy

Heterogeneous autonomous agents from different manufacturers cannot coordinate in shared physical spaces without either a central authority (which violates sovereignty) or full state sharing (which violates proprietary IP). Existing approaches—VDA 5050, ROS 2/Nav2, ORCA, and Dec-POMDPs—require homogeneity, centralization, or are computationally intractable for real-time physical coordination. We introduce Syncference, a five-phase coordination protocol where each agent projects a lossy five-field summary—the Minimum Viable Reality (MVR) —into a shared buffer. A conservative composition operator Γ (union of spatial claims, intersection of constraints, maximum of risk per cell) produces a collision-free coordination plan in sub-millisecond latency. Trust assumptions are minimal: agents do not reveal proprietary state, do not share a global objective, and do not reach consensus; a trust-degrading detection layer handles actively-adversarial participants. We validate Syncference empirically across two scenarios with N=500 runs per cell via MAZ3, a public Apache 2. 0 benchmark. Comparisons include ORCA (van den Berg et al. 2011) as a standard baseline and an OmniscientCoordinator that feeds lossless MVRs into the same composition operator Γ — isolating the cost of sovereignty as the single axis of comparison. All empirical hypotheses were submitted as a pre-registration to OSF on 2026-04-19 (DOI 10. 17605/OSF. IO/KJCWG, currently in OSF review queue) prior to the N=500 confirmatory run; all four were confirmed exactly. Scope note (critical): Syncference’s primary empirical value is demonstrated in geometrically saturated scenarios where alternative protocols exhibit collision rates of 0. 16–0. 27 per cycle (see Bottleneck results in §5) ; in geometrically benign scenarios with local myopic policies, simpler baselines like Greedy may achieve equivalent or superior task-level outcomes at lower coordination cost (see Asymmetric-Risk results in §5. 3). All task-level equivalence claims in this paper are conditional on the local myopic policy class tested (agents consult only the current-cell value of M) ; whether the observed Hₚ advantages of lossless composition translate to task-level differences under policies with bounded look-ahead is the explicit empirical target of planned follow-up work (§7. 2). Three headline findings: (a) Syncference achieves zero collisions across all 7, 500 Syncference runs; (b) under asymmetric-information conditions with local myopic policies, Γ with lossless inputs produces a composite M with strictly higher coordination coherence than Γ with lossy inputs (Hₚ 1. 000 vs 0. 65–0. 71, |Δ| = 0. 29–0. 35), but the difference does not propagate to task-level outcomes — a pre-registered directional negative prediction (H4) confirmed exactly; (c) an unexpected N=500 finding: ORCA under saturated corridor geometry exhibits 0. 22–0. 27 collisions per cycle with task completion ≤ 0. 05, a failure mode invisible at N=50 but emergent with larger samples. Additional exploratory findings on network-modulated composition advantage and adversarial detection are reported in §5. Syncference proposes a third paradigm for multi-entity coordination — distinct from both centralization and consensus — where conservative composition of partial, lossy projections produces a shared coordination surface without requiring objective sharing, full state disclosure, or global agreement. We release MAZ3 as a public benchmark and do not claim it is the only such benchmark; we claim it is a benchmark designed around sovereignty, adversariality, and documented failure modes.